Emotion Canvas Help

Find runosong poems by their emotional density — the share of words that belong to emotion families. Pick which emotion families to include, narrow by place, year or song type, and browse the matching poems sorted by density.

Using this page

• Family chips: Click a chip to toggle a single emotion family (joy, grief, love…). Click the domain header to toggle a whole group. Presets All / Positive / Negative / Clear quickly re-select common sets.
• Wordform panel: Lists every emotion wordform contained in the selected families — these are exactly what's counted in each poem's density score. Add a Place filter to see only the wordforms actually attested there.
• Filters (More filters): Year range, place, song type, minimum word count, minimum density, title search, and a single-wordform filter for finding poems that contain a specific form. Press / to jump to title search.
• Histogram: Click a bucket (0–1 %, 1–3 %, 3–5 %, 5–10 %, 10 %+) to keep only poems in that density range.
• Sort & browse: Sort the table by density, count, wordforms, year, verses, or diversity. Click any row to open the full poem with emotion words highlighted by family colour; use ◀ ▶ or arrow keys to page through matches.
• Export & share: Export CSV saves the current filtered list. The URL is a permalink to your selection (families + sort + filters), so it can be bookmarked or shared.

Where the emotion vocabulary comes from — 12 detection methods

The emotion families counted here are the same 38 families / 63,871 word forms as the Emotion Vocabulary browser. They were assembled by combining 12 detection methods. Most operate on English translations of lemmatized runosong texts (DeepSeek R1); the substitution test is the one method that works directly on Estonian verse templates.

• Lemmatization — EstNLTK + dialect corpus, Estonian and Finnish dictionaries, and LLM-based lemmatization (Pivovarova et al. 2025); four lemmatizers combined.
• Substitution test (Harris 1954) — each seed word is removed from its verse to form a template; every other word that fills the same slot elsewhere becomes a candidate. Templates are IDF-weighted; 26 emotion seeds were expanded to depth 5.
• LLM emotion extraction (Claude Opus) — filters substitution candidates into the E / M / A / J / N classes.
• SetFit (Tunstall et al. 2022) — fine-tuned text classifier that predicts emotion domains for all 1.2M translated forms.
• GoEmotions (Demszky et al. 2020) — RoBERTa classifier, 28 categories on each translation.
• NRC EmoLex — Plutchik scores for 14K English words, plus Estonian + Finnish NRC lexicons matched directly on source forms.
• DepecheMood++ — crowd-sourced emotion scores for 100K+ English lemmas, transferred via translations.
• GLiNER (Zaratiana et al. 2024) — zero-shot NER that filters out concrete entities (persons, animals, objects).
• WordNet (Princeton / NLTK) — blocks emotion labels on translations classified as animal / plant / food entities.
• Lemma-based expansion — once a lemma is confirmed, all its paradigmatic forms inherit the label; modern EKKD emotion dictionaries (v3–v7) are cross-referenced to validate expansions.

All methods are merged with consensus scoring (≥2 independent methods must agree). Each family is then reviewed by Claude Opus against verse evidence, and three layers of corrections are applied before the result propagates to the RunoVerse lexicon and this page's poem index.

Two different “substitution tests” on RunoVerse

The substitution test referenced above is the emotion-pipeline version: seed-driven, IDF-weighted, expanded to depth 5. The separate Substitution Explorer uses the main RunoVerse similarity pipeline — 701K cross-poem (XP) + parallelism (CP) pairs, not seed-driven, not restricted to emotions. The two datasets are complementary (only ~13.6% overlap).

Note: The vocabulary is computationally derived and AI-reviewed; estimated false-positive rate after consensus + Claude review is ~9%.

See also: Emotion Families for word-level browsing, and Emotional Geography for the parish-level map view.

?